In the early hours of October 9, Google Cloud published "Welcome to Gemini at Work 2026: Introducing the Gemini agent," a long post signed by CEO Thomas Kurian, officially unveiling Gemini Agent — a universal work agent for enterprises. Chinese financial media immediately covered it under the label "general-purpose work agent."

The product's working principle is a single sentence: you give it objectives, not instructions.Hand it a goal — say, a market analysis report — and it will research sources, build a financial model in Sheets, and assemble the deck in Slides on its own, choosing tools, planning steps, and spinning up sub-agents as needed, with event triggers and scheduled tasks for jobs that can run for days.

The "AI Coworker" and Four Layers of Memory

The most imaginative piece of the release is the Coworker: enterprises can give an agent a durable identity — its own Workspace account, corporate email, calendar and Drive storage — so it appears in the company directory, joins group chats and can be @-mentioned in documents. It acts under its own identity and leaves an audit trail.

To help this "colleague" improve with use, Google equipped it with four kinds of memory: session memory keeps context across multi-day tasks; semantic memory distills read documents and collaborations into a structured knowledge base; procedural memory records how a class of work gets done and can even save learned methods for reuse; episodic memory retains past execution history. Google's own metaphor: "like a new employee gradually getting to know you, your tools and your team."

Multi-Source Facts at a Glance

· Google states the agent can run on Gemini models or on Anthropic's Claude models (Cailianshe, Sina Tech).

· Multi-model orchestration with Smart Routing: cheap models for simple steps, stronger models for hard ones (QbitAI).

Swappable Foundation — Even to Claude

The counterintuitive part: Google lets enterprises pick the underlying model, including Claude from rival Anthropic. The rationale is pragmatic — model rankings keep shifting, and businesses should swap foundations without rebuilding workflows. The move shifts competition from single-model benchmarks to orchestration, data and governance layers, precisely where cloud vendors hold home advantage.

A Three-Way Agent War

The timing is delicate: Meta's Muse handles shopping, booking and payments for users; OpenAI's Dots gives agents a dedicated cloud computer running 24/7 across 4,000+ apps; and Gemini Agent is rooted in the Workspace ecosystem — a three-giant contest. For Chinese vendors the signal is clear: the next phase is not "can it finish tasks" but the combined depth of long-term memory, identity systems, security auditing and cost control.

NineZenith Research sees the same logic in enterprise deployments: the TianXing platform carries multi-agent orchestration and long-task scheduling, while the TianDun stack guards identity, permissions and audit trails — together they let "AI coworkers" enter production safely. (Information synthesized from QbitAI, Cailianshe, Jijie, GeekPark and other public reports)